Results and Discussion
10.1 Endmember extraction algorithms
10.1.1 Radarsat2 data
Results and Discussion
We shall now continue to discuss the most important and interesting results for all the experimental testes and simulations explained in the previous chapter.
This is in order to validate if the approach in this thesis works or not.
10.1 Endmember extraction algorithms
In this section we will cover and discus the results of using the different endmember extraction algorithms on the Radarsat2, the ALOS2 and the IceSAR dataset.
10.1.1 Radarsat2 data
In order to validate the performance of the endmember extraction algorithms on the Radarsat2 scene, the algorithms were tested on the ground truth. Table 10.1 to table 10.3 shows the ice thickness of the endmembers using the extraction algorithms on the ground truth data. The tables shows the results of using two to four endmembers.
The results from the ground truth dataset is quite poor as the extracted end-members had very similar ice thickness. Most of the endend-members were classified as ice with thickness between 0.8 meters to 2.2 meters. Figure 10.1 shows the
53
distribution of the ice thickness for the samples in the ground truth dataset. It also shows that there exists ice in the scene that are thicker then teen meters.
Good results from the extraction algorithms would resulted in endmembers with a larger difference in the ice thickness.
The bad results have probably been caused by the distribution of the data regarding the ice thickness. The ice thickness was gamma distributed. The samples which are lager then 3 meters, are few compared to those between 0.8 and 2.4 meter. Accordingly, these few samples becomes outliers and are located far away from the data manifold. The three endmember extraction algorithms will therefore have trouble locating all the pure pixels in the data.
If the distribution of the ice thickness was more uniform, the results would probably have been better.
Endmember nr. Ice thickness Endmember nr. Ice thickness Endmember nr. Ice thickness
1 2.1930 1 2.1930 1 2.1930
2 1.4920 2 1.0400 2 1.0400
3 0.6750 3 0.6750
4 1.4920
Table 10.1:Endmember classification. The table shows the ice thickness for the differ-ent endmembers extracted using PPI
Endmember nr. Ice thickness Endmember nr. Ice thickness Endmember nr. Ice thickness
1 2.1930 1 1.4920 1 1.0400
2 1.4920 2 2.1930 2 2.2940
3 0.8150 3 2.1930
4 0.6750
Table 10.2:Endmember classification. The table shows the ice thickness for the differ-ent endmembers extracted using NFINDR
Endmember nr. Ice thickness Endmember nr. Ice thickness Endmember nr. Ice thickness
1 1.4920 1 1.4920 1 1.4920
2 0.5270 2 0.5270 2 0.5270
3 2.1930 3 2.1930
4 0.6750
Table 10.3:Endmember classification. The table shows the ice thickness for the differ-ent endmembers extracted using ATGP
10.1 E N D M E M B E R E X T R AC T I O N A LG O R I T H M S 55
Figure 10.1: The distribution of the ice thickness in the ground truth dataset, x-axis is the ice thickness in meters and axis is the number of sample.
10.1.2 ALOS2
Figure 10.3, figure 10.5 and figure 10.7 shows the position of the extracted endmembers on the intensity image of the scene using the VV polarization. The endmembers are extracted from the ALOS2 data using the different endmember extraction algorithms. The figures shows the results of using two to four endmembers. One will note from the figures that the endmembers are well separated from each other. These results are possible to use in order to obtain a validation on the performance of the endmember extraction algorithms.
To receive the required validation discussed above, the corresponding Landsat8 data was used. Figure 10.4, figure 10.6 and figure 10.8 shows where the endmembers are located in the Landsat8 scene
By looking at figure 10.4 (a), figure 10.6 (a) and figure 10.8 (a), it is possible to note that all the endmember extraction algorithms detected different types of classes for all the three endmembers. All the algorithms had one endmember for open water, one endmember for thicker ice and one endmember for thinner ice.
In regards to figure 10.4 (b), figure 10.6 (b) and figure 10.8 (b), which shows the results using four endmembers, it is difficult to see what the fourth end-member is. The fourth endend-member for PPI appears as thicker ice, like the first PPI endmember. The situation is similar for the fourth endmember for both the NFINDR and ATGP algorithm. For the NFINDR algorithm, the fourth extracted endmember appears as thin ice, similar as the first endmember. The fourth extracted endmember using ATGP appears as thicker ice, like the first endmember. At this number of endmembers, this type of ground truth is limited in order to classifying the endmembers. Still, the ground truth indicates that the endmember extraction algorithms had success in finding different types of pure pixels.
However, even though the ground truth can’t tell if the four extracted endmem-bers are different from each other, it is still possible to answer the question. By calculating the Euclidean distance between the four extracted endmembers, it is possible to verify if the endmembers are different from each other. Table 10.4, 10.5 and 10.6 shows the results of calculating the Euclidean distance between the four extracted endmemers. The table 10.5 and 10.6 tells us that the four endmembers extracted using the NFINDR and ATGP algorithm are not similar, because the values in the tables are big and not small. However, the fourth endmember extracted by the PPI algorithm is similar to the first endmember, because the value is small. The smaller the values are, the more similar are the endmembers. Accordingly, when the value is zero the endmembers will be identical. One can also see this by studding figure 10.2, where endmember one and four have similar polarimetric values, resulting in similar shape on the plot.
Endmember 1 Endmember 2 Endmember 3 Endmember 4
Endmember 1 0 2.1552 2.2410 0.9108
Endmember 2 2.1552 0 2.2325 2.1138
Endmember 3 2.2410 2.2325 0 1.6761
Endmember 4 0.9108 2.1138 1.6761 0
Table 10.4:The table shows the Euclidean distance between the four endmemers extracted using the PPI algorithm
Endmember 1 Endmember 2 Endmember 3 Endmember 4
Endmember 1 0 2.2325 2.2410 1.6706
Endmember 2 2.2325 0 2.1552 1.3584
Endmember 3 2.2410 2.1552 0 2.0203
Endmember 4 1.6706 1.3584 2.0203 0
Table 10.5:The table shows the Euclidean distance between the four endmemers extracted using the NFINDR algorithm
10.1 E N D M E M B E R E X T R AC T I O N A LG O R I T H M S 57 Endmember 1 Endmember 2 Endmember 3 Endmember 4
Endmember 1 0 2.4110 2.1552 1.5375
Endmember 2 2.2410 0 2.2325 1.7401
Endmember 3 2.1552 2.2325 0 1.5191
Endmember 4 1.5375 1.74401 1.15191 0
Table 10.6:The table shows the Euclidean distance between the four endmemers extracted using the ATGP algorithm
Figure 10.2: The polarimetric values for the four endmembers extracted by PPI
(a) (b) (c)
Figure 10.3:The position in the VV intensity image for the two to four endmembers extracted using PPI algorithm. Figure (a) shows two extracted endmem-bers. Figure (b) shows three extracted endmemendmem-bers. Figure (c) four extracted endmembers.
(a) (b)
Figure 10.4: The corresponding position for the PPI endmembers in RGB image from the Landsat8 data. Figure (a) shows three extracted endmembers and figure (b) shows four extracted endmembers.
10.1 E N D M E M B E R E X T R AC T I O N A LG O R I T H M S 59
(a) (b) (c)
Figure 10.5:The position in the VV intensity image for the two to four endmembers extracted using NFINDR algorithm. Figure (a) shows two extracted endmembers. Figure (b) shows three extracted endmembers. Figure (c) four extracted endmembers.
(a) (b)
Figure 10.6: The corresponding position for the NFINDR endmembers in RGB image from the Landsat8 data. Figure (a) shows three extracted endmembers and figure (b) shows four extracted endmembers.
(a) (b) (c)
Figure 10.7:The position in the VV intensity image for the two to four endmembers extracted using ATGP algorithm. Figure (a) shows two extracted end-members. Figure (b) shows three extracted endend-members. Figure (c) four extracted endmembers.
(a) (b)
Figure 10.8: The corresponding position for the ATGP endmembers in RGB image from the Landsat8 data. Figure (a) shows three extracted endmembers and figure (b) shows four extracted endmembers.
10.1 E N D M E M B E R E X T R AC T I O N A LG O R I T H M S 61